Evaluation metrics for synthetic medical imaging.

In 30 seconds

This systematic review evaluated metrics for assessing synthetic medical imaging across 47 studies. It found significant variability in evaluation practices, with no single metric effectively capturing fidelity, realism, diversity, and clinical validity. The most common metrics included peak signal-to-noise ratio (34%) and structural similarity index (32%).

Key findings

• Expert evaluation and reference-based evaluations were the most common methods, each used in 53% of studies.
• Peak signal-to-noise ratio (PSNR) was the most frequently used metric, appearing in 34% of studies.
• No single metric effectively assessed all dimensions of synthetic medical images simultaneously.

Why it matters

The lack of standardized evaluation metrics for synthetic medical imaging hinders reproducibility and comparability across studies, which is crucial for clinical application and trust in AI-generated images.